EEG stress classification
NeuroSentiScan
An offline, interpretable comparison of classical and deep models for three-state stress classification from engineered EEG features.
EEG-derived signals are noisy, high-dimensional and personal. NeuroSentiScan compares model families for distinguishing stressed, normal and relaxed states while examining which features influence predictions.
Which classical and deep-learning approaches best separate three cognitive states in the available engineered EEG dataset?
My contribution
- Contributed to the framework and manuscript as the second listed author.
- Worked across classical/deep model comparisons, feature analysis, confusion matrices and explainability.
- Supported statistical comparison using paired tests and interpretation with LIME/SHAP.
Signature view
Trace the evidence.
Move across the research chain. The case study below remains complete and readable without this interaction.
Question
Separate three cognitive states from EEG features
The goal is a careful comparison, not only a single best-model headline.
Method, as a sequence of decisions
- 01
A balanced 2,132-sample dataset represents stressed, normal and relaxed classes.
- 02
Statistical, entropy and frequency-domain descriptors form 2,548 engineered features.
- 03
Six classical and six deep architectures are compared on consistent stratified splits.
- 04
Feature importance, LIME/SHAP and statistical tests complement aggregate metrics.
What the record actually supports
Samples in manuscript
708 stressed, 716 normal and 708 relaxed samples.
Engineered features
Statistical, entropy-based and frequency-domain inputs.
Reported CNN-1D accuracy
A manuscript result on the available split, not an external or cross-subject validation.
Evidence: inspected NeuroSentiScan manuscript. Status follows the latest profile: manuscript in preparation. Reported metrics are not external validation.
Choices and trade-offs
Why compare classical models?
Strong engineered features can let smaller models compete while remaining cheaper to train and easier to inspect.
Why add explanations?
Aggregate accuracy cannot show whether a model is using plausible spectral and nonlinear descriptors or exploiting an artefact.
What has not been established
- The manuscript remains in preparation; no verified publication or conference presentation is claimed.
- The reported split does not establish cross-subject or real-time generalisation.
- High-dimensional engineered features can amplify leakage or subject-specific structure if splitting is not carefully controlled.
What I would investigate next
- Q1How does leave-one-subject-out validation change the ranking of models?
- Q2Can temporal models operate directly on minimally processed signals without losing interpretability?
- Q3Which explanation patterns replicate across subjects and recording conditions?
Peer notes / NeuroSentiScan
See a missed assumption?
Research suggestions, methodological critique and relevant collaboration ideas are particularly useful.